AI Data Training Jobs

AI Data Training Jobs — independent reviews, comparisons, pricing and step-by-step guides on Aizhi.

• 

  SAP Cloud Infrastructure

  SAP Cloud Infrastructure is an SAP-operated IaaS cloud platform, used to run SAP’s cloud business and customer-facing deployments for SAP and non-SAP workloads. It is developed and operated with open-source technologies within SAP’s data center network, based on OpenStack and Kubernetes and supporting SAP S/4HANA and general-purpose applications. It offers compute, storage, and platform services that are accessible to SAP customers. == History == In 2012, SAP promoted aspects of cloud computing. In October 2012, SAP announced a platform as a service called the SAP Cloud Platform. In May 2013, a managed private cloud called the S/4HANA Enterprise Cloud service was announced. SAP Converged Cloud was announced in January 2015. SAP Converged Cloud was originally developed as SAP's internal standardized Infrastructure as a Service (IaaS) offering to support SAP’s cloud solutions. Originating from SAP Converged Cloud, SAP Cloud Infrastructure was developed and announced as SAP’s cloud computing offering that is provided for both SAP and customer workloads. In 2025, it had a global footprint of 15 regions and 29 data centers, encompassing more than 200,000 active VMs and over 6,000 hypervisors. In September 2025, SAP announced an expansion of its European “SAP Sovereign Cloud” portfolio, explicitly naming SAP Cloud Infrastructure (alongside SAP Sovereign Cloud On-Site) as part of the stack positioned for public sector and regulated environments. == Services and Features == SAP Cloud Infrastructure (SCI) is an infrastructure-as-a-service (IaaS) offering by SAP that provides virtual compute, storage, and networking services, together with identity, key management, and operational services. SCI follows a self-service model and is managed via APIs and a web-based user interface. === Compute === SCI provides virtual machine instances that can be provisioned from operating system images and selected in predefined sizes (“flavors”). It supports lifecycle operations such as create/modify/resize/delete, power control, and snapshots; instances can be organized into server groups to influence placement policies. === Storage === SCI provides persistent storage services including: Block storage (virtual volumes) with attach/detach to instances, online expansion, cloning, snapshots, and provisioning volumes from images or snapshots. Object storage (containers and objects) managed via API/CLI with access control lists (ACLs) and configurable redundancy options. File storage (shared file systems) with access controls, online resize, snapshots/restore, and replication across availability zones. === Networking === SCI provides software-defined networking (SDN) for tenant networks (networks, subnets, routers) and connectivity features such as floating IPs for public reachability. Network security controls include security groups and firewall policies; connectivity options include BGP-based VPN networking. === Load balancing and DNS === SCI includes managed load balancing for distributing traffic across backend instances and an authoritative DNS service (DNSaaS) with API-based management of DNS zones and records, including options for zone sharing/transfer across projects/tenants and service integrations for automated record creation. === Identity, access, and key management === SCI includes identity and access management for authentication/authorization in projects/tenants (for example token handling, role assignment, and credential management) and key/secrets management for storing and controlling access to secret material such as keys and certificates, including support for different backends (depending on configuration). === Cloud-native services === SCI includes a container image registry (image push/pull, access policies, and lifecycle controls) and an auto-scaling capability for file shares based on configurable rules. === Observability and audit === SCI includes metrics and audit logging capabilities for operational monitoring and for listing/filtering audit-relevant events across services. === Availability and service levels === SCI documentation describes availability-related features such as load balancing, storage redundancy options, and replication for file shares across availability zones. SAP cloud services are governed by contractual service-level agreements (SLA); SAP Cloud Infrastructure references an SLA supplement defining infrastructure-specific terms when referenced in order forms. === SAP cloud services === SAP cloud services can run on different underlying infrastructures, including SAP Cloud Infrastructure in addition to SAP NS2 or hyperscalers. SAP cloud solutions available on SAP Cloud Infrastructure include SAP Cloud ERP, SAP HCM, SAP Solutions for Spend Management, Supply Chain Management, Business Transformation Management, and SAP Business Technology Platform (including related analytics and business data solutions). For example, SAP HANA Cloud documentation lists SAP Cloud Infrastructure as one of the supported infrastructures alongside hyperscalers. === Sustainability === SAP describes sustainability initiatives for its data centers, including energy-efficient infrastructure (for example, advanced cooling systems and power management), renewable electricity usage where feasible, and operational practices such as recycling electronic waste and minimizing water usage. SAP also references environmental management and energy management standards such as ISO 14001 and ISO 50001 for its data center operations. SAP-owned data centers run with 100% renewable electricity and that renewable electricity has been used since 2014 to power SAP facilities including owned data centers and co-locations. == SAP Cloud Infrastructure for SAP Sovereign Cloud == SAP Sovereign Cloud is a portfolio of SAP solutions designed to help organizations adopt SAP cloud solutions such as the SAP Cloud ERP while maintaining control over data, infrastructure, and compliance in line with local laws and regulations. The portfolio offers multiple deployment options, including SAP Cloud Infrastructure and SAP Sovereign Cloud On-Site, alongside sovereign hyperscaler-based options such as via SAP NS2, and targets customers such as public-sector bodies and other highly regulated organizations. In Europe, SAP Cloud Infrastructure is an Infrastructure-as-a-Service (IaaS) deployment option within SAP Sovereign Cloud for SAP and customer / third party workloads, operated on SAP’s data center network and developed using open-source technologies, with customer data stored within the European Union. Sovereignty-related characteristics for the SAP Cloud Infrastructure include: EU footprint and ownership model: SAP-operated data centers in Germany include sites in St. Leon-Rot and Walldorf, and co-location sites in Frankfurt. EU AI Cloud: EU AI Cloud is a sovereign AI offering for Europe that provides secure, compliant environments for building and running AI, including governed access to auditable large language models from SAP and partners. It offers AI models on the SAP Cloud Infrastructure and SAP Business Technology Platform (SAP BTP), enabling deployment of AI applications and models on high-performance European infrastructure (including accelerator/GPU-based compute for AI workloads). Availability zones and secure interconnect: Three availability zones in three independent data centers in Germany, connected via SAP-owned fiber on SAP-owned property. Facility and security standards: ISO/IEC 27001 governance of delivery and operations of SAP cloud services and SAP-owned data centers. Additional facility and availability standards: EN 50600 availability class 3 (European data centre standard) and/or ISO/IEC 22237 availability class 3 (international equivalent). Technology foundation: Based on open-source cloud infrastructure framework (OpenStack) and Kubernetes, without dependencies on hyperscaler technologies. Sovereignty controls: Data sovereignty (data residency), operational sovereignty (administration and maintenance restricted to approved, security-cleared personnel), technical sovereignty (locally hosted control planes with separation via encryption or dedicated infrastructure), and legal sovereignty (use of locally based legal entities or those in approved countries). Classified information processing: Roadmap to meet high and very high requirements for handling classified or sensitive information under European regulatory and security regimes. Public-sector readiness and EU sovereignty assurance levels: Implemented to meet SEAL-3 (Digital Resilience) and SEAL-4 (Full Digital Sovereignty) of the European Commission’s Cloud Sovereignty Framework. Staffing constraints: Operations model selectable to restrict sensitive operations to vetted personnel from EU or NATO countries.

  Read more →
• 

  Arabic Speech Corpus

  The Arabic Speech Corpus is a Modern Standard Arabic (MSA) speech corpus for speech synthesis. The corpus contains phonetic and orthographic transcriptions of more than 3.7 hours of MSA speech aligned with recorded speech on the phoneme level. The annotations include word stress marks on the individual phonemes. The Arabic Speech Corpus was built as part of a doctoral project by Nawar Halabi at the University of Southampton funded by MicroLinkPC who own an exclusive license to commercialise the corpus, but the corpus is available for strictly non-commercial purposes through the official Arabic Speech Corpus website. It is distributed under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License. == Purpose == The corpus was mainly built for speech synthesis purposes, specifically Speech Synthesis, but the corpus has been used for building HMM based voices in Arabic. It was also used to automatically align other speech corpora with their phonetic transcript and could be used as part of a larger corpus for training speech recognition systems. == Contents == The package contains the following: 1813 .wav files containing spoken utterances. 1813 .lab files containing text utterances. 1813 .TextGrid files containing the phoneme labels with time stamps of the boundaries where these occur in the .wav files. phonetic-transcript.txt which has the form "[wav_filename]" "[Phoneme Sequence]" in every line. orthographic-transcript.txt which has the form "[wav_filename]" "[Orthographic Transcript]" in every line. Orthography is in Buckwalter Format which is friendlier where there is software that does not read Arabic script. It can be easily converted back to Arabic. There is an extra 18 minutes of fully annotated corpus (separate from above but with the same structure as above) which was used to evaluated the corpus (see PhD thesis). The corpus was also used to prove that using automatically extracted, orthography-based stress marks improve the quality of speech synthesis in MSA.

  Read more →
• 

  Averaged one-dependence estimators

  Averaged one-dependence estimators (AODE) is a probabilistic classification learning technique. It was developed to address the attribute-independence problem of the popular naive Bayes classifier. It frequently develops substantially more accurate classifiers than naive Bayes at the cost of a modest increase in the amount of computation. == The AODE classifier == AODE seeks to estimate the probability of each class y given a specified set of features x1, ... xn, P(y | x1, ... xn). To do so it uses the formula P ^ ( y ∣ x 1 , … x n ) = ∑ i : 1 ≤ i ≤ n ∧ F ( x i ) ≥ m P ^ ( y , x i ) ∏ j = 1 n P ^ ( x j ∣ y , x i ) ∑ y ′ ∈ Y ∑ i : 1 ≤ i ≤ n ∧ F ( x i ) ≥ m P ^ ( y ′ , x i ) ∏ j = 1 n P ^ ( x j ∣ y ′ , x i ) {\displaystyle {\hat {P}}(y\mid x_{1},\ldots x_{n})={\frac {\sum _{i:1\leq i\leq n\wedge F(x_{i})\geq m}{\hat {P}}(y,x_{i})\prod _{j=1}^{n}{\hat {P}}(x_{j}\mid y,x_{i})}{\sum _{y^{\prime }\in Y}\sum _{i:1\leq i\leq n\wedge F(x_{i})\geq m}{\hat {P}}(y^{\prime },x_{i})\prod _{j=1}^{n}{\hat {P}}(x_{j}\mid y^{\prime },x_{i})}}} where P ^ ( ⋅ ) {\displaystyle {\hat {P}}(\cdot )} denotes an estimate of P ( ⋅ ) {\displaystyle P(\cdot )} , F ( ⋅ ) {\displaystyle F(\cdot )} is the frequency with which the argument appears in the sample data and m is a user specified minimum frequency with which a term must appear in order to be used in the outer summation. In recent practice m is usually set at 1. == Derivation of the AODE classifier == We seek to estimate P(y | x1, ... xn). By the definition of conditional probability P ( y ∣ x 1 , … x n ) = P ( y , x 1 , … x n ) P ( x 1 , … x n ) . {\displaystyle P(y\mid x_{1},\ldots x_{n})={\frac {P(y,x_{1},\ldots x_{n})}{P(x_{1},\ldots x_{n})}}.} For any 1 ≤ i ≤ n {\displaystyle 1\leq i\leq n} , P ( y , x 1 , … x n ) = P ( y , x i ) P ( x 1 , … x n ∣ y , x i ) . {\displaystyle P(y,x_{1},\ldots x_{n})=P(y,x_{i})P(x_{1},\ldots x_{n}\mid y,x_{i}).} Under an assumption that x1, ... xn are independent given y and xi, it follows that P ( y , x 1 , … x n ) = P ( y , x i ) ∏ j = 1 n P ( x j ∣ y , x i ) . {\displaystyle P(y,x_{1},\ldots x_{n})=P(y,x_{i})\prod _{j=1}^{n}P(x_{j}\mid y,x_{i}).} This formula defines a special form of One Dependence Estimator (ODE), a variant of the naive Bayes classifier that makes the above independence assumption that is weaker (and hence potentially less harmful) than the naive Bayes' independence assumption. In consequence, each ODE should create a less biased estimator than naive Bayes. However, because the base probability estimates are each conditioned by two variables rather than one, they are formed from less data (the training examples that satisfy both variables) and hence are likely to have more variance. AODE reduces this variance by averaging the estimates of all such ODEs. == Features of the AODE classifier == Like naive Bayes, AODE does not perform model selection and does not use tuneable parameters. As a result, it has low variance. It supports incremental learning whereby the classifier can be updated efficiently with information from new examples as they become available. It predicts class probabilities rather than simply predicting a single class, allowing the user to determine the confidence with which each classification can be made. Its probabilistic model can directly handle situations where some data are missing. AODE has computational complexity O ( l n 2 ) {\displaystyle O(ln^{2})} at training time and O ( k n 2 ) {\displaystyle O(kn^{2})} at classification time, where n is the number of features, l is the number of training examples and k is the number of classes. This makes it infeasible for application to high-dimensional data. However, within that limitation, it is linear with respect to the number of training examples and hence can efficiently process large numbers of training examples. == Implementations == The free Weka machine learning suite includes an implementation of AODE.

  Read more →
• 

  Alternating decision tree

  An alternating decision tree (ADTree) is a machine learning method for classification. It generalizes decision trees and has connections to boosting. An ADTree consists of an alternation of decision nodes, which specify a predicate condition, and prediction nodes, which contain a single number. An instance is classified by an ADTree by following all paths for which all decision nodes are true, and summing any prediction nodes that are traversed. == History == ADTrees were introduced by Yoav Freund and Llew Mason. However, the algorithm as presented had several typographical errors. Clarifications and optimizations were later presented by Bernhard Pfahringer, Geoffrey Holmes and Richard Kirkby. Implementations are available in Weka and JBoost. == Motivation == Original boosting algorithms typically used either decision stumps or decision trees as weak hypotheses. As an example, boosting decision stumps creates a set of T {\displaystyle T} weighted decision stumps (where T {\displaystyle T} is the number of boosting iterations), which then vote on the final classification according to their weights. Individual decision stumps are weighted according to their ability to classify the data. Boosting a simple learner results in an unstructured set of T {\displaystyle T} hypotheses, making it difficult to infer correlations between attributes. Alternating decision trees introduce structure to the set of hypotheses by requiring that they build off a hypothesis that was produced in an earlier iteration. The resulting set of hypotheses can be visualized in a tree based on the relationship between a hypothesis and its "parent." Another important feature of boosted algorithms is that the data is given a different distribution at each iteration. Instances that are misclassified are given a larger weight while accurately classified instances are given reduced weight. == Alternating decision tree structure == An alternating decision tree consists of decision nodes and prediction nodes. Decision nodes specify a predicate condition. Prediction nodes contain a single number. ADTrees always have prediction nodes as both root and leaves. An instance is classified by an ADTree by following all paths for which all decision nodes are true and summing any prediction nodes that are traversed. This is different from binary classification trees such as CART (Classification and regression tree) or C4.5 in which an instance follows only one path through the tree. === Example === The following tree was constructed using JBoost on the spambase dataset (available from the UCI Machine Learning Repository). In this example, spam is coded as 1 and regular email is coded as −1. The following table contains part of the information for a single instance. The instance is scored by summing all of the prediction nodes through which it passes. In the case of the instance above, the score is calculated as The final score of 0.657 is positive, so the instance is classified as spam. The magnitude of the value is a measure of confidence in the prediction. The original authors list three potential levels of interpretation for the set of attributes identified by an ADTree: Individual nodes can be evaluated for their own predictive ability. Sets of nodes on the same path may be interpreted as having a joint effect The tree can be interpreted as a whole. Care must be taken when interpreting individual nodes as the scores reflect a re weighting of the data in each iteration. == Description of the algorithm == The inputs to the alternating decision tree algorithm are: A set of inputs ( x 1 , y 1 ) , … , ( x m , y m ) {\displaystyle (x_{1},y_{1}),\ldots ,(x_{m},y_{m})} where x i {\displaystyle x_{i}} is a vector of attributes and y i {\displaystyle y_{i}} is either -1 or 1. Inputs are also called instances. A set of weights w i {\displaystyle w_{i}} corresponding to each instance. The fundamental element of the ADTree algorithm is the rule. A single rule consists of a precondition, a condition, and two scores. A condition is a predicate of the form "attribute value." A precondition is simply a logical conjunction of conditions. Evaluation of a rule involves a pair of nested if statements: 1 if (precondition) 2 if (condition) 3 return score_one 4 else 5 return score_two 6 end if 7 else 8 return 0 9 end if Several auxiliary functions are also required by the algorithm: W + ( c ) {\displaystyle W_{+}(c)} returns the sum of the weights of all positively labeled examples that satisfy predicate c {\displaystyle c} W − ( c ) {\displaystyle W_{-}(c)} returns the sum of the weights of all negatively labeled examples that satisfy predicate c {\displaystyle c} W ( c ) = W + ( c ) + W − ( c ) {\displaystyle W(c)=W_{+}(c)+W_{-}(c)} returns the sum of the weights of all examples that satisfy predicate c {\displaystyle c} The algorithm is as follows: 1 function ad_tree 2 input Set of m training instances 3 4 wi = 1/m for all i 5 a = 1 2 ln W + ( t r u e ) W − ( t r u e ) {\displaystyle a={\frac {1}{2}}{\textrm {ln}}{\frac {W_{+}(true)}{W_{-}(true)}}} 6 R0 = a rule with scores a and 0, precondition "true" and condition "true." 7 P = { t r u e } {\displaystyle {\mathcal {P}}=\{true\}} 8 C = {\displaystyle {\mathcal {C}}=} the set of all possible conditions 9 for j = 1 … T {\displaystyle j=1\dots T} 10 p ∈ P , c ∈ C {\displaystyle p\in {\mathcal {P}},c\in {\mathcal {C}}} get values that minimize z = 2 ( W + ( p ∧ c ) W − ( p ∧ c ) + W + ( p ∧ ¬ c ) W − ( p ∧ ¬ c ) ) + W ( ¬ p ) {\displaystyle z=2\left({\sqrt {W_{+}(p\wedge c)W_{-}(p\wedge c)}}+{\sqrt {W_{+}(p\wedge \neg c)W_{-}(p\wedge \neg c)}}\right)+W(\neg p)} 11 P + = p ∧ c + p ∧ ¬ c {\displaystyle {\mathcal {P}}+=p\wedge c+p\wedge \neg c} 12 a 1 = 1 2 ln W + ( p ∧ c ) + 1 W − ( p ∧ c ) + 1 {\displaystyle a_{1}={\frac {1}{2}}{\textrm {ln}}{\frac {W_{+}(p\wedge c)+1}{W_{-}(p\wedge c)+1}}} 13 a 2 = 1 2 ln W + ( p ∧ ¬ c ) + 1 W − ( p ∧ ¬ c ) + 1 {\displaystyle a_{2}={\frac {1}{2}}{\textrm {ln}}{\frac {W_{+}(p\wedge \neg c)+1}{W_{-}(p\wedge \neg c)+1}}} 14 Rj = new rule with precondition p, condition c, and weights a1 and a2 15 w i = w i e − y i R j ( x i ) {\displaystyle w_{i}=w_{i}e^{-y_{i}R_{j}(x_{i})}} 16 end for 17 return set of Rj The set P {\displaystyle {\mathcal {P}}} grows by two preconditions in each iteration, and it is possible to derive the tree structure of a set of rules by making note of the precondition that is used in each successive rule. == Empirical results == Figure 6 in the original paper demonstrates that ADTrees are typically as robust as boosted decision trees and boosted decision stumps. Typically, equivalent accuracy can be achieved with a much simpler tree structure than recursive partitioning algorithms.

  Read more →
• 

  AI content watermarking

  AI content watermarking is the process of embedding imperceptible yet detectable signals into content generated by artificial intelligence systems, such as text, images, audio, or video. The technique allows the content to be traced and identified as machine-generated without compromising its quality for the end user. AI watermarking has emerged as a key approach to address growing concerns about misinformation, deepfakes, copyright infringement, and the traceability of synthetic content in the context of the rapid development of generative artificial intelligence. Unlike traditional visible watermarks used in photography, AI content watermarks are typically invisible to humans and can only be detected and deciphered algorithmically. The concept is distinct from the watermarking of AI models themselves (to prevent model theft) and from the watermarking of training data (to combat unauthorized data use). Modern AI watermarking schemes are typically formalized as a pair of algorithms, an embedding (or generation) algorithm and a detection algorithm, sharing a secret key, whose performance is evaluated along three competing axes: quality (the watermark must not noticeably degrade outputs), detectability (the watermark must be statistically distinguishable from unwatermarked content), and robustness (the watermark must persist under adversarial or incidental modifications). == Background == Digital watermarking has been used for decades to protect physical and digital media, from paper currency to photographs. Classical schemes typically embedded a fixed bit-string into a fixed cover signal, with robustness criteria defined against a small fixed set of distortions such as JPEG compression or additive Gaussian noise. The rapid advancement of generative AI in the early 2020s, however, created a new and qualitatively different demand: rather than protecting a single artifact, watermarks for AI content must be embedded automatically across an open-ended distribution of generated outputs while remaining robust to a much wider class of adversarial transformations, including paraphrasing, image regeneration via diffusion models, and re-recording. Large image generation models such as DALL-E, Stable Diffusion, and Midjourney, along with large language models like ChatGPT, made it possible to produce highly realistic synthetic text, images, audio, and video at scale, raising significant ethical and security concerns. In July 2023, the Biden administration secured voluntary commitments from leading AI companies, including OpenAI, Alphabet, Meta, and Amazon, to develop watermarking and other provenance technologies to help users identify AI-generated content. == Formal definitions and design goals == Most modern AI watermarking schemes can be formalized as a pair of algorithms ( W m , D e t e c t ) {\displaystyle ({\mathsf {Wm}},{\mathsf {Detect}})} parameterized by a secret key k {\displaystyle k} . The embedding algorithm W m {\displaystyle {\mathsf {Wm}}} takes a generative model M {\displaystyle M} (and optionally a prompt) and returns a watermarked output x {\displaystyle x} ; the detection algorithm D e t e c t ( x , k ) {\displaystyle {\mathsf {Detect}}(x,k)} outputs a real-valued score (typically a p-value or log-likelihood ratio) used to decide whether x {\displaystyle x} was produced by the watermarked generator. The literature evaluates such schemes along several largely conflicting criteria: Criteria for evaluation include imperceptibility or quality preservation, measured for text via perplexity and human preference judgments, and for images and audio via metrics such as PSNR, SSIM, LPIPS, or PESQ. Detectability is typically expressed as the true positive rate at a fixed false positive rate (e.g. 1% or 10^-6), or as the number of tokens or pixels needed to reach a given confidence level. Robustness refers to the requirement that the watermark should survive expected modifications like JPEG or MP3 compression, cropping, noise, paraphrasing, or machine translation. Distortion-freeness is a stronger property requiring that the marginal distribution of any single watermarked output be statistically identical to the unwatermarked model's distribution. Schemes due to Aaronson, Christ et al., and Kuditipudi et al. are distortion-free in this sense, while the original Kirchenbauer et al. scheme is not. Forgery resistance or unforgeability means an adversary without the secret key should be unable to produce content that passes detection. == Techniques == AI watermarking techniques vary significantly depending on the type of content being watermarked. At its core, the process involves two main stages: embedding (or encoding) the watermark, and detection. There are two primary methods for embedding: watermarking during content generation, which requires access to the AI model itself but is generally more robust, and post-generation watermarking, which can be applied to content from any source, including closed-source models. Watermarks can be broadly classified as visible, including overt marks such as logos or text overlays, or imperceptible, which are detectable only by algorithms. They can also be classified by durability: robust watermarks are designed to withstand common transformations such as compression, cropping, and re-encoding, while fragile watermarks are easily destroyed by any alteration, making them useful for tamper detection. A further axis distinguishes zero-bit watermarks, which only signal "this content was generated by model M," from multi-bit watermarks, which embed an arbitrary payload (such as a user identifier) that can be recovered at detection time. === Text === Text watermarking is considered one of the most challenging modalities because natural language offers relatively limited redundancy compared to images or audio. Modern approaches for large language models alter the autoregressive sampling process so that some statistical signature is left in the choice of tokens, while leaving the surface form of the text unchanged. The literature distinguishes three main families of generation-time text watermarks. Logit-biasing schemes (e.g. KGW) add a fixed bias δ {\displaystyle \delta } to a pseudorandomly selected subset of vocabulary logits before softmax sampling. Reweighting or sampling-based schemes (e.g. SynthID-Text) compose multiple pseudorandom tournaments over the model's full distribution. Distortion-free schemes based on the Gumbel-max trick or inverse transform sampling (Aaronson 2022; Kuditipudi et al. 2023; Christ et al. 2024) preserve the marginal output distribution of the model. ==== KGW: token-probability shifting ==== The pioneering "green list / red list" scheme of Kirchenbauer et al. (KGW), introduced at ICML 2023, is the foundation for most subsequent text watermarks. At each decoding step t {\displaystyle t} , a pseudorandom function (PRF) keyed by a secret k {\displaystyle k} is applied to a context window of h {\displaystyle h} previous tokens to deterministically partition the vocabulary V {\displaystyle V} of size N {\displaystyle N} into a "green list" G ⊂ V {\displaystyle G\subset V} of size γ N {\displaystyle \gamma N} and its complement, the "red list" R = V ∖ G {\displaystyle R=V\setminus G} , where γ ∈ ( 0 , 1 ) {\displaystyle \gamma \in (0,1)} (typically γ = 1 / 2 {\displaystyle \gamma =1/2} ) is the green fraction. A logits processor then increments every green-list logit by a fixed bias δ > 0 {\displaystyle \delta >0} before softmax: ℓ v ′ = ℓ v + δ ⋅ 1 [ v ∈ G ] {\displaystyle \ell '_{v}=\ell _{v}+\delta \cdot \mathbf {1} [v\in G]} so that, after sampling, green tokens are over-represented but generation is not constrained to green tokens alone; high-entropy positions tolerate the bias gracefully, while low-entropy positions (where one token dominates the logits) override the watermark and preserve correctness on factual content. Detection requires only the secret key and the candidate text, not the language model itself. The detector recomputes the partition g ( ⋅ ) {\displaystyle g(\cdot )} for each token, counts the number of green hits | G | hits {\displaystyle |G|_{\text{hits}}} in a sequence of length T {\displaystyle T} , and computes a one-proportion z-test statistic: z = | G | hits − γ T T γ ( 1 − γ ) {\displaystyle z={\frac {|G|_{\text{hits}}-\gamma T}{\sqrt {T\gamma (1-\gamma )}}}} Under the null hypothesis that the text was written by an unwatermarked source (human or another model), the green-hit count is approximately binomially distributed with mean γ T {\displaystyle \gamma T} ; a large positive z {\displaystyle z} rejects the null hypothesis. The original paper reports that fewer than 25 watermarked tokens are sufficient to detect a watermark with a false positive rate below 10^-5 on the OPT-1.3B model. A follow-up study by the same group documented robustness under temperature sampling, top-p (nucleus) sampling, and human paraphrasing, and proposed sliding-window

  Read more →
• 

  Non-negative matrix factorization

  Non-negative matrix factorization (NMF or NNMF), also non-negative matrix approximation is a group of algorithms in multivariate analysis and linear algebra where a matrix V is factorized into (usually) two matrices W and H, with the property that all three matrices have no negative elements. This non-negativity makes the resulting matrices easier to inspect. Also, in applications such as processing of audio spectrograms or muscular activity, non-negativity is inherent to the data being considered. Since the problem is not exactly solvable in general, it is commonly approximated numerically. NMF finds applications in such fields as astronomy, computer vision, document clustering, missing data imputation, chemometrics, audio signal processing, recommender systems, and bioinformatics. == History == In chemometrics non-negative matrix factorization has a long history under the name "self modeling curve resolution". In this framework the vectors in the right matrix are continuous curves rather than discrete vectors. Also early work on non-negative matrix factorizations was performed by a Finnish group of researchers in the 1990s under the name positive matrix factorization. It became more widely known as non-negative matrix factorization after Lee and Seung investigated the properties of the algorithm and published some simple and useful algorithms for two types of factorizations. == Background == Let matrix V be the product of the matrices W and H, V = W H . {\displaystyle \mathbf {V} =\mathbf {W} \mathbf {H} \,.} Matrix multiplication can be implemented as computing the column vectors of V as linear combinations of the column vectors in W using coefficients supplied by columns of H. That is, each column of V can be computed as follows: v i = W h i , {\displaystyle \mathbf {v} _{i}=\mathbf {W} \mathbf {h} _{i}\,,} where vi is the i-th column vector of the product matrix V and hi is the i-th column vector of the matrix H. When multiplying matrices, the dimensions of the factor matrices may be significantly lower than those of the product matrix and it is this property that forms the basis of NMF. NMF generates factors with significantly reduced dimensions compared to the original matrix. For example, if V is an m × n matrix, W is an m × p matrix, and H is a p × n matrix then p can be significantly less than both m and n. Here is an example based on a text-mining application: Let the input matrix (the matrix to be factored) be V with 10000 rows and 500 columns where words are in rows and documents are in columns. That is, we have 500 documents indexed by 10000 words. It follows that a column vector v in V represents a document. Assume we ask the algorithm to find 10 features in order to generate a features matrix W with 10000 rows and 10 columns and a coefficients matrix H with 10 rows and 500 columns. The product of W and H is a matrix with 10000 rows and 500 columns, the same shape as the input matrix V and, if the factorization worked, it is a reasonable approximation to the input matrix V. From the treatment of matrix multiplication above it follows that each column in the product matrix WH is a linear combination of the 10 column vectors in the features matrix W with coefficients supplied by the coefficients matrix H. This last point is the basis of NMF because we can consider each original document in our example as being built from a small set of hidden features. NMF generates these features. It is useful to think of each feature (column vector) in the features matrix W as a document archetype comprising a set of words where each word's cell value defines the word's rank in the feature: The higher a word's cell value the higher the word's rank in the feature. A column in the coefficients matrix H represents an original document with a cell value defining the document's rank for a feature. We can now reconstruct a document (column vector) from our input matrix by a linear combination of our features (column vectors in W) where each feature is weighted by the feature's cell value from the document's column in H. == Clustering property == NMF has an inherent clustering property, i.e., it automatically clusters the columns of input data V = ( v 1 , … , v n ) {\displaystyle \mathbf {V} =(v_{1},\dots ,v_{n})} . More specifically, the approximation of V {\displaystyle \mathbf {V} } by V ≃ W H {\displaystyle \mathbf {V} \simeq \mathbf {W} \mathbf {H} } is achieved by finding W {\displaystyle W} and H {\displaystyle H} that minimize the error function (using the Frobenius norm) ‖ V − W H ‖ F , {\displaystyle \left\|V-WH\right\|_{F},} subject to W ≥ 0 , H ≥ 0. {\displaystyle W\geq 0,H\geq 0.} , If we furthermore impose an orthogonality constraint on H {\displaystyle \mathbf {H} } , i.e. H H T = I {\displaystyle \mathbf {H} \mathbf {H} ^{T}=I} , then the above minimization is mathematically equivalent to the minimization of K-means clustering. Furthermore, the computed H {\displaystyle H} gives the cluster membership, i.e., if H k j > H i j {\displaystyle \mathbf {H} _{kj}>\mathbf {H} _{ij}} for all i ≠ k, this suggests that the input data v j {\displaystyle v_{j}} belongs to k {\displaystyle k} -th cluster. The computed W {\displaystyle W} gives the cluster centroids, i.e., the k {\displaystyle k} -th column gives the cluster centroid of k {\displaystyle k} -th cluster. This centroid's representation can be significantly enhanced by convex NMF. When the orthogonality constraint H H T = I {\displaystyle \mathbf {H} \mathbf {H} ^{T}=I} is not explicitly imposed, the orthogonality holds to a large extent, and the clustering property holds too. When the error function to be used is Kullback–Leibler divergence, NMF is identical to the probabilistic latent semantic analysis (PLSA), a popular document clustering method. == Types == === Approximate non-negative matrix factorization === Usually the number of columns of W and the number of rows of H in NMF are selected so the product WH will become an approximation to V. The full decomposition of V then amounts to the two non-negative matrices W and H as well as a residual U, such that: V = WH + U. The elements of the residual matrix can either be negative or positive. When W and H are smaller than V they become easier to store and manipulate. Another reason for factorizing V into smaller matrices W and H, is that if one's goal is to approximately represent the elements of V by significantly less data, then one has to infer some latent structure in the data. === Convex non-negative matrix factorization === In standard NMF, matrix factor W ∈ R+m × k， i.e., W can be anything in that space. Convex NMF restricts the columns of W to convex combinations of the input data vectors ( v 1 , … , v n ) {\displaystyle (v_{1},\dots ,v_{n})} . This greatly improves the quality of data representation of W. Furthermore, the resulting matrix factor H becomes more sparse and orthogonal. === Nonnegative rank factorization === In case the nonnegative rank of V is equal to its actual rank, V = WH is called a nonnegative rank factorization (NRF). The problem of finding the NRF of V, if it exists, is known to be NP-hard. === Different cost functions and regularizations === There are different types of non-negative matrix factorizations. The different types arise from using different cost functions for measuring the divergence between V and WH and possibly by regularization of the W and/or H matrices. Two simple divergence functions studied by Lee and Seung are the squared error (or Frobenius norm) and an extension of the Kullback–Leibler divergence to positive matrices (the original Kullback–Leibler divergence is defined on probability distributions). Each divergence leads to a different NMF algorithm, usually minimizing the divergence using iterative update rules. The factorization problem in the squared error version of NMF may be stated as: Given a matrix V {\displaystyle \mathbf {V} } find nonnegative matrices W and H that minimize the function F ( W , H ) = ‖ V − W H ‖ F 2 {\displaystyle F(\mathbf {W} ,\mathbf {H} )=\left\|\mathbf {V} -\mathbf {WH} \right\|_{F}^{2}} Another type of NMF for images is based on the total variation norm. When L1 regularization (akin to Lasso) is added to NMF with the mean squared error cost function, the resulting problem may be called non-negative sparse coding due to the similarity to the sparse coding problem, although it may also still be referred to as NMF. === Online NMF === Many standard NMF algorithms analyze all the data together; i.e., the whole matrix is available from the start. This may be unsatisfactory in applications where there are too many data to fit into memory or where the data are provided in streaming fashion. One such use is for collaborative filtering in recommendation systems, where there may be many users and many items to recommend, and it would be inefficient to recalculate everything when one user or one item is added to the system. The cost function for o

  Read more →
• 

  Universal approximation theorem

  In the field of machine learning, the universal approximation theorems (UATs) state that neural networks with a certain structure can, in principle, approximate any continuous function to any desired degree of accuracy. These theorems provide a mathematical justification for using neural networks, assuring researchers that a sufficiently large or deep network can model the complex, non-linear relationships often found in real-world data. The best-known version of the theorem applies to feedforward networks with a single hidden layer. It states that if the layer's activation function is non-polynomial (which is true for common choices like the sigmoid function or ReLU), then the network can act as a "universal approximator." Universality is achieved by increasing the number of neurons in the hidden layer, making the network "wider." Other versions of the theorem show that universality can also be achieved by keeping the network's width fixed but increasing its number of layers, making it "deeper." These are existence theorems. They guarantee that a network with the right structure exists, but they do not provide a method for finding the network's parameters (training it), nor do they specify exactly how large the network must be for a given function. Finding a suitable network remains a practical challenge that is typically addressed with optimization algorithms like backpropagation. == Setup == Artificial neural networks are combinations of multiple simple mathematical functions that implement more complicated functions from (typically) real-valued vectors to real-valued vectors. The spaces of multivariate functions that can be implemented by a network are determined by the structure of the network, the set of simple functions, and its multiplicative parameters. A great deal of theoretical work has gone into characterizing these function spaces. Most universal approximation theorems are in one of two classes. The first quantifies the approximation capabilities of neural networks with an arbitrary number of artificial neurons ("arbitrary width" case) and the second focuses on the case with an arbitrary number of hidden layers, each containing a limited number of artificial neurons ("arbitrary depth" case). In addition to these two classes, there are also universal approximation theorems for neural networks with bounded number of hidden layers and a limited number of neurons in each layer ("bounded depth and bounded width" case). == History == === Arbitrary width === The first results concerned the arbitrary width case. Ken-ichi Funahashi (May 1989) showed that Rumelhart–Hinton–Williams type backpropagation networks possess universal approximation capability with a class of sigmoidal activation functions, extending the result to multi-output mappings as well. Kurt Hornik, Maxwell Stinchcombe, and Halbert White (July 1989) showed that multilayer feed-forward networks with as few as one hidden layer are universal approximators, provided that the activation function satisfies certain conditions. George Cybenko (December 1989) independently established a related result for sigmoid activation functions using functional-analytic methods. Hornik also showed in 1991 that it is not the specific choice of the activation function but rather the multilayer feed-forward architecture itself that gives neural networks the potential of being universal approximators. Moshe Leshno et al in 1993 and later Allan Pinkus in 1999 showed that the universal approximation property is equivalent to having a nonpolynomial activation function. === Arbitrary depth === The arbitrary depth case was also studied by a number of authors such as Gustaf Gripenberg in 2003, Dmitry Yarotsky, Zhou Lu et al in 2017, Boris Hanin and Mark Sellke in 2018 who focused on neural networks with ReLU activation function. In 2020, Patrick Kidger and Terry Lyons extended those results to neural networks with general activation functions such, e.g. tanh or GeLU. One special case of arbitrary depth is that each composition component comes from a finite set of mappings. In 2024, Cai constructed a finite set of mappings, named a vocabulary, such that any continuous function can be approximated by compositing a sequence from the vocabulary. This is similar to the concept of compositionality in linguistics, which is the idea that a finite vocabulary of basic elements can be combined via grammar to express an infinite range of meanings. === Bounded depth and bounded width === The bounded depth and bounded width case was first studied by Maiorov and Pinkus in 1999. They showed that there exists an analytic sigmoidal activation function such that two hidden layer neural networks with bounded number of units in hidden layers are universal approximators. In 2018, Guliyev and Ismailov constructed a smooth sigmoidal activation function providing universal approximation property for two hidden layer feedforward neural networks with fewer units in hidden layers. In 2018, they also constructed single hidden layer networks with bounded width that are still universal approximators for univariate functions. However, this does not apply for multivariable functions. In 2022, Shen et al. obtained precise quantitative information on the depth and width required to approximate a target function by deep and wide ReLU neural networks. === Quantitative bounds === The question of minimal possible width for universality was first studied in 2021, Park et al obtained the minimum width required for the universal approximation of Lp functions using feed-forward neural networks with ReLU as activation functions. Similar results that can be directly applied to residual neural networks were also obtained in the same year by Paulo Tabuada and Bahman Gharesifard using control-theoretic arguments. In 2023, Cai obtained the optimal minimum width bound for the universal approximation. For the arbitrary depth case, Leonie Papon and Anastasis Kratsios derived explicit depth estimates depending on the regularity of the target function and of the activation function. === Kolmogorov network === The Kolmogorov–Arnold representation theorem is similar in spirit. Indeed, certain neural network families can directly apply the Kolmogorov–Arnold theorem to yield a universal approximation theorem. Robert Hecht-Nielsen showed that a three-layer neural network can approximate any continuous multivariate function. This was extended to the discontinuous case by Vugar Ismailov. In 2024, Ziming Liu and co-authors showed a practical application. === Reservoir computing and quantum reservoir computing === In reservoir computing a sparse recurrent neural network with fixed weights equipped of fading memory and echo state property is followed by a trainable output layer. Its universality has been demonstrated separately for what concerns networks of rate neurons and spiking neurons, respectively. In 2024, the framework has been generalized and extended to quantum reservoirs where the reservoir is based on qubits defined over Hilbert spaces. === Variants === Variants include discontinuous activation functions, noncompact domains, certifiable networks, random neural networks, and alternative network architectures and topologies. The universal approximation property of width-bounded networks has been studied as a dual of classical universal approximation results on depth-bounded networks. For input dimension d x {\displaystyle d_{x}} and output dimension d y {\displaystyle d_{y}} the minimum width required for the universal approximation of the Lp functions is exactly m a x { d x + 1 , d y } {\displaystyle max\{d_{x}+1,d_{y}\}} (for a ReLU network). More generally this also holds if both ReLU and a threshold activation function are used. Universal function approximation on graphs (or rather on graph isomorphism classes) by popular graph convolutional neural networks (GCNs or GNNs) can be made as discriminative as the Weisfeiler–Leman graph isomorphism test. In 2020, a universal approximation theorem result was established by Brüel-Gabrielsson, showing that graph representation with certain injective properties is sufficient for universal function approximation on bounded graphs and restricted universal function approximation on unbounded graphs, with an accompanying O ( | V | ⋅ | E | ) {\displaystyle {\mathcal {O}}(\left|V\right|\cdot \left|E\right|)} -runtime method that performed at state of the art on a collection of benchmarks (where V {\displaystyle V} and E {\displaystyle E} are the sets of nodes and edges of the graph respectively). There are also a variety of results between non-Euclidean spaces and other commonly used architectures and, more generally, algorithmically generated sets of functions, such as the convolutional neural network (CNN) architecture, radial basis functions, or neural networks with specific properties. == Arbitrary-width case == A universal approximation theorem formally states that a family of neural network funct

  Read more →
• 

  Q-learning

  Q-learning is a reinforcement learning algorithm that trains an agent to assign values to its possible actions based on its current state, without requiring a model of the environment (model-free). It can handle problems with stochastic transitions and rewards without requiring adaptations. For example, in a grid maze, an agent learns to reach an exit worth 10 points. At a junction, Q-learning might assign a higher value to moving right than left if right gets to the exit faster, improving this choice by trying both directions over time. For any finite Markov decision process, Q-learning finds an optimal policy in the sense of maximizing the expected value of the total reward over any and all successive steps, starting from the current state. Q-learning can identify an optimal action-selection policy for any given finite Markov decision process, given infinite exploration time and a partly random policy. "Q" refers to the function that the algorithm computes: the expected reward—that is, the quality—of an action taken in a given state. == Reinforcement learning == Reinforcement learning involves an agent, a set of states S {\displaystyle {\mathcal {S}}} , and a set A {\displaystyle {\mathcal {A}}} of actions per state. By performing an action a ∈ A {\displaystyle a\in {\mathcal {A}}} , the agent transitions from state to state. Executing an action in a specific state provides the agent with a reward (a numerical score). The goal of the agent is to maximize its total reward. It does this by adding the maximum reward attainable from future states to the reward for achieving its current state, effectively influencing the current action by the potential future reward. This potential reward is a weighted sum of expected values of the rewards of all future steps starting from the current state. As an example, consider the process of boarding a train, in which the reward is measured by the negative of the total time spent boarding (alternatively, the cost of boarding the train is equal to the boarding time). One strategy is to enter the train door as soon as they open, minimizing the initial wait time for yourself. If the train is crowded, however, then you will have a slow entry after the initial action of entering the door as people are fighting you to depart the train as you attempt to board. The total boarding time, or cost, is then: 0 seconds wait time + 15 seconds fight time On the next day, by random chance (exploration), you decide to wait and let other people depart first. This initially results in a longer wait time. However, less time is spent fighting the departing passengers. Overall, this path has a higher reward than that of the previous day, since the total boarding time is now: 5 second wait time + 0 second fight time Through exploration, despite the initial (patient) action resulting in a larger cost (or negative reward) than in the forceful strategy, the overall cost is lower, thus revealing a more rewarding strategy. == Algorithm == After Δ t {\displaystyle \Delta t} steps into the future the agent will decide some next step. The weight for this step is calculated as γ Δ t {\displaystyle \gamma ^{\Delta t}} , where γ {\displaystyle \gamma } (the discount factor) is a number between 0 and 1 ( 0 ≤ γ ≤ 1 {\displaystyle 0\leq \gamma \leq 1} ). Assuming γ < 1 {\displaystyle \gamma <1} , it has the effect of valuing rewards received earlier higher than those received later (reflecting the value of a "good start"). γ {\displaystyle \gamma } may also be interpreted as the probability to succeed (or survive) at every step Δ t {\displaystyle \Delta t} . The algorithm, therefore, has a function that calculates the quality of a state–action combination: Q : S × A → R {\displaystyle Q:{\mathcal {S}}\times {\mathcal {A}}\to \mathbb {R} } . Before learning begins, ⁠ Q {\displaystyle Q} ⁠ is initialized to a possibly arbitrary fixed value (chosen by the programmer). Then, at each time t {\displaystyle t} the agent selects an action A t {\displaystyle A_{t}} , observes a reward R t + 1 {\displaystyle R_{t+1}} , enters a new state S t + 1 {\displaystyle S_{t+1}} (that may depend on both the previous state S t {\displaystyle S_{t}} and the selected action), and Q {\displaystyle Q} is updated. The core of the algorithm is a Bellman equation as a simple value iteration update, using the weighted average of the current value and the new information: Q n e w ( S t , A t ) ← ( 1 − α ⏟ learning rate ) ⋅ Q ( S t , A t ) ⏟ current value + α ⏟ learning rate ⋅ ( R t + 1 ⏟ reward + γ ⏟ discount factor ⋅ max a Q ( S t + 1 , a ) ⏟ estimate of optimal future value ⏟ new value (temporal difference target) ) {\displaystyle Q^{new}(S_{t},A_{t})\leftarrow (1-\underbrace {\alpha } _{\text{learning rate}})\cdot \underbrace {Q(S_{t},A_{t})} _{\text{current value}}+\underbrace {\alpha } _{\text{learning rate}}\cdot {\bigg (}\underbrace {\underbrace {R_{t+1}} _{\text{reward}}+\underbrace {\gamma } _{\text{discount factor}}\cdot \underbrace {\max _{a}Q(S_{t+1},a)} _{\text{estimate of optimal future value}}} _{\text{new value (temporal difference target)}}{\bigg )}} where R t + 1 {\displaystyle R_{t+1}} is the reward received when moving from the state S t {\displaystyle S_{t}} to the state S t + 1 {\displaystyle S_{t+1}} , and α {\displaystyle \alpha } is the learning rate ( 0 < α ≤ 1 ) {\displaystyle (0<\alpha \leq 1)} . Note that Q n e w ( S t , A t ) {\displaystyle Q^{new}(S_{t},A_{t})} is the sum of three terms: ( 1 − α ) Q ( S t , A t ) {\displaystyle (1-\alpha )Q(S_{t},A_{t})} : the current value (weighted by one minus the learning rate) α R t + 1 {\displaystyle \alpha \,R_{t+1}} : the reward R t + 1 {\displaystyle R_{t+1}} to obtain if action A t {\displaystyle A_{t}} is taken when in state S t {\displaystyle S_{t}} (weighted by learning rate) α γ max a Q ( S t + 1 , a ) {\displaystyle \alpha \gamma \max _{a}Q(S_{t+1},a)} : the maximum reward that can be obtained from state S t + 1 {\displaystyle S_{t+1}} (weighted by learning rate and discount factor) An episode of the algorithm ends when state S t + 1 {\displaystyle S_{t+1}} is a final or terminal state. However, Q-learning can also learn in non-episodic tasks (as a result of the property of convergent infinite series). If the discount factor is lower than 1, the action values are finite even if the problem can contain infinite loops or paths. For all final states s f {\displaystyle s_{f}} , Q ( s f , a ) {\displaystyle Q(s_{f},a)} is never updated, but is set to the reward value r {\displaystyle r} observed for state s f {\displaystyle s_{f}} . In most cases, Q ( s f , a ) {\displaystyle Q(s_{f},a)} can be taken to equal zero. == Influence of variables == === Learning rate === The learning rate or step size determines to what extent newly acquired information overrides old information. A factor of 0 makes the agent learn nothing (exclusively exploiting prior knowledge), while a factor of 1 makes the agent consider only the most recent information (ignoring prior knowledge to explore possibilities). In fully deterministic environments, a learning rate of α t = 1 {\displaystyle \alpha _{t}=1} is optimal. When the problem is stochastic, the algorithm converges under some technical conditions on the learning rate that require it to decrease to zero. In practice, often a constant learning rate is used, such as α t = 0.1 {\displaystyle \alpha _{t}=0.1} for all t {\displaystyle t} . === Discount factor === The discount factor ⁠ γ {\displaystyle \gamma } ⁠ determines the importance of future rewards. A factor of 0 will make the agent "myopic" (or short-sighted) by only considering current rewards, i.e. r t {\displaystyle r_{t}} (in the update rule above), while a factor approaching 1 will make it strive for a long-term high reward. If the discount factor meets or exceeds 1, the action values may diverge. For ⁠ γ = 1 {\displaystyle \gamma =1} ⁠, without a terminal state, or if the agent never reaches one, all environment histories become infinitely long, and utilities with additive, undiscounted rewards generally become infinite. Even with a discount factor only slightly lower than 1, Q-function learning leads to propagation of errors and instabilities when the value function is approximated with an artificial neural network. In that case, starting with a lower discount factor and increasing it towards its final value accelerates learning. === Initial conditions (Q0) === Since Q-learning is an iterative algorithm, it implicitly assumes an initial condition before the first update occurs. High initial values, also known as "optimistic initial conditions", can encourage exploration: no matter what action is selected, the update rule will cause it to have lower values than the other alternative, thus increasing their choice probability. The first reward r {\displaystyle r} can be used to reset the initial conditions. According to this idea, the first time an action is taken the reward is used to set the value

  Read more →
• 

  Webmail

  Webmail (or web-based email) is an email service that can be accessed using a standard web browser. It contrasts with email service accessible through a specialised email client software. Additionally, many internet service providers (ISP) provide webmail as part of their internet service package. Similarly, some web hosting providers also provide webmail as a part of their hosting package. As with any web application, webmail's main advantage over the use of a desktop email client is the ability to send and receive email anywhere from a web browser. == History == === Early implementations === The first Web Mail implementation was developed at CERN in 1993 by Phillip Hallam-Baker as a test of the HTTP protocol stack, but was not developed further. In the next two years, however, several people produced working webmail applications. In Europe, there were three implementations, Søren Vejrum's "WWW Mail", Luca Manunza's "WebMail", and Remy Wetzels' "WebMail". Søren Vejrum's "WWW Mail" was written when he was studying and working at the Copenhagen Business School in Denmark, and was released on February 28, 1995. Luca Manunza's "WebMail" was written while he was working at CRS4 in Sardinia, from an idea of Gianluigi Zanetti, with the first source release on March 30, 1995. Remy Wetzels' "WebMail" was written while he was studying at the Eindhoven University of Technology in the Netherlands for the DSE and was released early January 1995. In the United States, Matt Mankins wrote "Webex", and Bill Fitler, while at Lotus cc:Mail, began working on an implementation which he demonstrated publicly at Lotusphere on January 24, 1995. Customers who saw the cc:Mail demonstration were very enthusiastic, one recalling that they were "like an angry mob. People were yelling, 'We want this now!'". Matt Mankins, under the supervision of Dr. Burt Rosenberg at the University of Miami, released his "Webex" application source code in a post to comp.mail.misc on August 8, 1995, although it had been in use as the primary email application at the School of Architecture where Mankins worked for some months prior. Bill Fitler's webmail implementation was further developed as a commercial product, which Lotus announced and released in the fall of 1995 as cc:Mail for the World Wide Web 1.0; thereby providing an alternative means of accessing a cc:Mail message store (the usual means being a cc:Mail desktop application that operated either via dialup or within the confines of a local area network). Early commercialization of webmail was also achieved when "Webex" began to be sold by Mankins' company, DotShop, Inc., at the end of 1995. Within DotShop, "Webex" changed its name to "EMUmail"; which would be sold to companies like UPS and Rackspace until its sale to Accurev in 2001. EMUmail was one of the first applications to feature a free version that included embedded advertising, as well as a licensed version that did not. Hotmail and Four11's RocketMail both launched in 1996 as free services and immediately became very popular. === Widespread deployment === As the 1990s progressed, and into the 2000s, it became more common for the general public to have access to webmail because: many Internet service providers (such as EarthLink) and web hosting providers (such as Verio) began bundling webmail into their service offerings (often in parallel with POP/SMTP services); many other enterprises (such as universities and large corporations) also started offering webmail as a way for their user communities to access their email (either locally managed or outsourced); webmail service providers (such as Hotmail and RocketMail) emerged in 1996 as a free service to the general public, and rapidly gained in popularity. In some cases, webmail application software is developed in-house by the organizations running and managing the application, and in some cases it is obtained from software companies that develop and sell such applications, usually as part of an integrated mail server package (an early example being Netscape Messaging Server). The market for webmail application software has continued into the 2010s. == Rendering and compatibility == Email users may find the use of both a webmail client and a desktop client using the POP3 protocol presents some difficulties. For example, email messages that are downloaded by the desktop client and are removed from the server will no longer be available on the webmail client. The user is limited to previewing messages using the web client before they are downloaded by the desktop email client. However, one may choose to leave the emails on the server, in which case this problem does not occur. The use of both a webmail client and a desktop client using the IMAP4 protocol allows the contents of the mailbox to be consistently displayed in both the webmail and desktop clients and any action the user performs on messages in one interface will be reflected when the email is accessed via the other interface. There are significant differences in rendering capabilities for many popular webmail services such as Gmail, Outlook.com and Yahoo! Mail. Due to the varying treatment of HTML tags, such as